lluten/

ptb_pcfg

Synthetic English pretraining corpora generated from a Penn Treebank-derived constituency PCFG.

This release contains three corpus variants with different subject-verb agreement cue reliability settings:

rel_0.5 rel_0.75 rel_1.0

Each variant contains:

en/tokens.txt en/tokens.txt.metadata.json

The corpora use opaque token IDs rather than English word forms and were created for experiments on structural transfer from… See the full description on the dataset page: https://huggingface.co/datasets/lluten/ptb_pcfg.